@nicholas Replied by email. Thank you for reaching out, Nicholas.

@mrinallovesbhature sent you an inquiry to the email on your website.

@Ridwan

Hey, I make pro bono software for educational/research agencies. Let me know if you need pro bono software that has features that are normally out of budget. I have clients in your region (e.g., Papua New Guinea, India)

Best Regards,
Nicholas Volta
nicholas@voltarecovery.org
WhatsApp: (willing to send privately)

Progress update, day one: the codebase is now public. Goal 4 from the plan above is done, eleven months ahead of schedule. Seiche is open source under AGPL 3.0 at https://github.com/LiquiLensLabs/seiche, one clean release commit with the full engine suite and the roughly 200 automated tests included. The funding manifest is live at seiche.info/funding.json and the project is now listed in the FLOSS/fund directory as well. The terminal kept publishing through all of it. Today's forecasts are sealed in the ledger as usual.

@Ridwan https://tinyurl.com/2djfzfe2

If you want to check my claims rather than take my word for it: the board is live at seiche.info, updates twice a day. Scroll to the bottom of today's letter, the section called "The record, nothing hidden." That's the backtest with the misses listed by name, not just the wins. All the data is free and keyless (FRED, NY Fed, OFR, Treasury, CFTC, ECB) so you can rebuild any number yourself if you're feeling suspicious. Honestly the most interesting thing to poke at is the ML calibration failure. The model ranks risk fine but its probabilities were junk, and the terminal says so on the page. Ask me anything, including the awkward stuff. I'd rather lose the grant than dodge a good question.

The DCB Technical Appendix (PoC) is now available for review. It includes 30-run Monte Carlo simulation results, performance comparisons, stability analysis, and energy efficiency calculations.

Link: https://drive.google.com/file/d/1NRW3KcnMb7rodkL7m8ffU5bV0ebRNVEA/view?usp=drivesdk

Feel free to ask questions or request additional details. I will update this thread as the project progresses.

For transparency: the noise-decomposition study referenced above was conducted on a separate memory-system evaluation (ES-MemEval). It is included in the repository because it empirically establishes DriftBench's premise — LLM-judge pipelines carry measurable, style-correlated noise — which the deterministic protocol addresses.

Progress update (July 9). Published an evaluation-reliability showcase in the repository (eval_reliability/): a measured three-layer noise decomposition on ES-MemEval (WWW '26) — blind human relabel of the LLM judge (K=20: overall noise 0.10, but per-arm leniency up to +0.33 that reorders the leaderboard), answerer stability (±0.05 on byte-identical inputs), and ingest stochasticity dominating subset-level scores (±0.40 swings between two ingests; format ablation contributed 0.00). Both pre-registrations are included, with recorded opposing predictions that the ablations later adjudicated. Directly relevant to DriftBench's thesis: deterministic evaluation exists because LLM-judge pipelines carry measurable, style-correlated noise. Also newly submitted: Anthropic External Researcher application (API credits to scale the judge study to K≈200 and a multi-judge panel).

What progress have you made since your last update?

Since the last update, I have completed the BLAM Liberia seed pilot and produced the main project report. The project helped turn a broad problem - Liberia’s weak biosafety and biosecurity capacity identified in the 2023 JEE into a more practical assessment framework. I developed and piloted the Biosafety and Biosecurity Landscape Assessment Matrix (BLAM) to look at concrete laboratory-system gaps, including incident reporting, pathogen accountability, waste management, equipment maintenance, specimen referral, training, and governance.

The pilot confirmed that the problem is real, but also more nuanced than “lack of resources.” Liberia has strong public health leadership, committed laboratory staff, and useful institutional networks. The bigger challenge is converting those strengths into routine systems: clear standards, regular supervision, maintenance, reporting, and follow up.

The main output is now a BLAM scoping report and a clearer pathway for future work. The pilot also helped me refine the project direction. I now see BLAM not only as a biosafety assessment tool, but as a platform that can later be extended to test whether fragile laboratory systems are ready to use AI safely in biological work. 

What are your next steps?

My next step is to use the pilot report to seek follow-on support for a focused scale-up phase. The immediate plan is to refine BLAM into a cleaner, easier-to-use tool with a scoring guide, indicators, and practical reporting templates. I also want to develop a small AI-biosecurity add-on module that asks a simple question: can public health laboratories safely use AI tools for tasks such as drafting SOPs, summarizing guidance, or preparing reports without exposing sensitive information or trusting incorrect advice?

The next phase would likely focus on the National Public Health Reference Laboratory in Monrovia and a small number of linked county laboratories. It would test BLAM more systematically, run a simple AI-biosecurity tabletop exercise, and produce practical outputs such as a safe-use guide, verification checklist, and escalation pathway for AI-related laboratory mistakes. Longer term, I want BLAM to become a Liberia-owned tool that can support better biosafety investment, partner coordination, and preparedness planning.

Is there anything others could help you with?

First, technical feedback from people with experience in biosafety, biosecurity, laboratory systems, or AI-biosecurity would be very useful. I especially need reviewers who can help make BLAM simpler, safer, and more credible before wider use.

Second, introductions to potential funders or collaborators interested in global health security, AI safety, pandemic preparedness, or laboratory strengthening would help move the work from seed pilot to scale up.

Third, I would appreciate practical support in turning the BLAM report into stronger public-facing materials: a short concept note, a clean tool/manual, and a funder-ready scale-up proposal. My PhD is currently unfunded, so support for protected time is also important if I am to keep pushing this work forward properly.

What progress have you made since your last update?

We re-analyzed the pilot data and quantified the headline effect: a small oracle budget raised the dishonest debater's win rate by +7.2pp (95% CI [4.6, 10.2]).

Before scaling up, we audited the pilot code and found two serious bugs in the oracle channel: every "NOT ADDRESSED" reply was miscoded to "NO", and ~100% of oracle queries were sent garbled. We retracted the mechanism conclusions and corrected the write-ups.

We then rebuilt the harness and launched a pre-registered validation run (in flight now, ~$150-200 of the grant): the same 318 transcripts re-judged under six arms, including a clean harness, a faithful bug replay, and a placebo oracle. Gates were frozen before any clean data existed. Pre-registration: https://github.com/jackmaiorino/selvarath-debate/blob/rerun-new-models/docs/rejudge-protocol.md and corrected report: https://github.com/jackmaiorino/selvarath-debate/blob/rerun-new-models/reports/2026-07-06-preliminary-findings.md

What are your next steps?

The pre-registered gates decide: if the effect survives the clean harness, we proceed to the proposed judge x debater capability grid (most of the grant). If the placebo explains it, we pivot to studying deliberation effects. If it collapses, we publish the artifact result and a decomposition of what each bug contributed. Write-up either way, including negative results.

Is there anything others could help you with?

Methods scrutiny of the pre-registered protocol before results land, and pointers to related work on oracle/verification interfaces or deliberation-length effects in LLM judging.

@JueYan Is it rude to ask you to take a look at my project? I'm bootstrapping since over a year and it's slowing things down by now. Thanks if you have/take the time to look, it's all I ask (apart from the funding ofc ;)

@nicholas Thank You SO very much!

That's a very generous offer, thank you @nicholas

I'll send you a mail tomorrow :)

@Lisa-Intel I think you missed the point haha, I'm providing the software to clients for no charge. I am accepting donations, though. E-mail me at nicholas@voltarecovery.org if you want to discuss this further.

Hi Nicholas, thanks!

I'd say my biggest bottleneck is building the core platform fast enough while staying focused. There are definitely areas where additional software could accelerate development, especially around AI orchestration, governance, integrations and developer tooling. Right now I'm keeping external spending very low while raising funding, but I'd love to stay in touch.

Coincidentally, I'm also exploring potential collaborations with researchers at the University of Florida (Department of Electrical and Computer Engineering) around autonomous AI systems, so it's nice to connect with you too.
@nicholas

Hey Pedro, out of curiosity, except for money, what is your biggest pain points that can be made into a software however it will cost a lot to create? I'm looking for clients honestly.

What i am trying to reach :

AATM started as a fundamental research project: can artificial cognition emerge from architecture, rather than from explicitly programmed cognitive functions?

The project has now found a major application axis: embodied intelligence for robotics.

One of the hardest problems in this field is not perception.
It is not movement.
It is not even learning.

It is internal regulation.

In living organisms, homeostasis is not a simple configuration file. The stability of internal states is the result of billions of years of evolution. Trying to manually find the perfect internal balance for one artificial organism is likely to be fragile, slow and structurally risky.

AATM changes the scaling strategy.

The goal is not to build one perfect artificial organism.

The goal is to build populations of imperfect blobs.

Each blob is a small autonomous sensorimotor unit: it has its own internal state, its own perception loop, its own movement, its own prediction errors, its own reinforcement dynamics and its own mini mammalian-inspired brain.

Some blobs may become too exploratory.
Some may become too conservative.
Some may collapse.
Some may become apathetic.
Some may discover useful behavioral patterns.

That is not a failure of the architecture.

That is the tuning surface.

Instead of fine-tuning one monolithic agent from the inside, AATM introduces a higher layer above the population. This layer can observe blobs over time, compare their behaviors, modify their parameters, duplicate useful configurations, remove unstable ones and consolidate what emerges across the swarm.

The higher layer does not need to have the same homeostasis as the blobs.

The blobs are driven by local regulation: action, energy, tension, prediction, novelty, saturation.

The upper layer is driven by population-level regulation: diversity, stability, coverage, robustness and collective performance.

This is the key shift.

AATM does not try to solve embodied intelligence by putting everything into one body and hoping the internal state remains stable.

It turns the problem into a distributed architecture:

local embodied agents below,
structural consolidation above.

This makes failure local instead of global.
It allows fine-tuning without restarting the whole system.
It creates a path from digital demonstrators to robotic swarms.
And it gives embodied AI a scalable architecture beyond the monolithic model.

What started as fundamental research into artificial cognition is becoming a practical path toward robotizable artificial organisms: imperfect by design, tunable by population, and scalable through collective consolidation.

Note: This grant was given as part of the experimental alignment microgrant program. Because this is a weird experimental grant program aiming to make lots of unusual bets, grantees are not allowed to use this grant as a credential when applying to any other grant, or job, or on their resume/linkedin/etc. This page exists because Manifund does not allow private grants. If you are a grantmaker or hiring manager, please disregard this grant when making decisions.

Maya, I appreciate the trust. Would you like to have a call? I also very much care about wild animals. Also happy to explain anything or keep you up to date on things

@nicholas Replied by email. Thank you for reaching out, Nicholas.

@mrinallovesbhature sent you an inquiry to the email on your website.

@Ridwan

Hey, I make pro bono software for educational/research agencies. Let me know if you need pro bono software that has features that are normally out of budget. I have clients in your region (e.g., Papua New Guinea, India)

Best Regards,
Nicholas Volta
nicholas@voltarecovery.org
WhatsApp: (willing to send privately)

Progress update, day one: the codebase is now public. Goal 4 from the plan above is done, eleven months ahead of schedule. Seiche is open source under AGPL 3.0 at https://github.com/LiquiLensLabs/seiche, one clean release commit with the full engine suite and the roughly 200 automated tests included. The funding manifest is live at seiche.info/funding.json and the project is now listed in the FLOSS/fund directory as well. The terminal kept publishing through all of it. Today's forecasts are sealed in the ledger as usual.

@Ridwan https://tinyurl.com/2djfzfe2

If you want to check my claims rather than take my word for it: the board is live at seiche.info, updates twice a day. Scroll to the bottom of today's letter, the section called "The record, nothing hidden." That's the backtest with the misses listed by name, not just the wins. All the data is free and keyless (FRED, NY Fed, OFR, Treasury, CFTC, ECB) so you can rebuild any number yourself if you're feeling suspicious. Honestly the most interesting thing to poke at is the ML calibration failure. The model ranks risk fine but its probabilities were junk, and the terminal says so on the page. Ask me anything, including the awkward stuff. I'd rather lose the grant than dodge a good question.

The DCB Technical Appendix (PoC) is now available for review. It includes 30-run Monte Carlo simulation results, performance comparisons, stability analysis, and energy efficiency calculations.

Link: https://drive.google.com/file/d/1NRW3KcnMb7rodkL7m8ffU5bV0ebRNVEA/view?usp=drivesdk

Feel free to ask questions or request additional details. I will update this thread as the project progresses.

For transparency: the noise-decomposition study referenced above was conducted on a separate memory-system evaluation (ES-MemEval). It is included in the repository because it empirically establishes DriftBench's premise — LLM-judge pipelines carry measurable, style-correlated noise — which the deterministic protocol addresses.

Progress update (July 9). Published an evaluation-reliability showcase in the repository (eval_reliability/): a measured three-layer noise decomposition on ES-MemEval (WWW '26) — blind human relabel of the LLM judge (K=20: overall noise 0.10, but per-arm leniency up to +0.33 that reorders the leaderboard), answerer stability (±0.05 on byte-identical inputs), and ingest stochasticity dominating subset-level scores (±0.40 swings between two ingests; format ablation contributed 0.00). Both pre-registrations are included, with recorded opposing predictions that the ablations later adjudicated. Directly relevant to DriftBench's thesis: deterministic evaluation exists because LLM-judge pipelines carry measurable, style-correlated noise. Also newly submitted: Anthropic External Researcher application (API credits to scale the judge study to K≈200 and a multi-judge panel).

What progress have you made since your last update?

Since the last update, I have completed the BLAM Liberia seed pilot and produced the main project report. The project helped turn a broad problem - Liberia’s weak biosafety and biosecurity capacity identified in the 2023 JEE into a more practical assessment framework. I developed and piloted the Biosafety and Biosecurity Landscape Assessment Matrix (BLAM) to look at concrete laboratory-system gaps, including incident reporting, pathogen accountability, waste management, equipment maintenance, specimen referral, training, and governance.

The pilot confirmed that the problem is real, but also more nuanced than “lack of resources.” Liberia has strong public health leadership, committed laboratory staff, and useful institutional networks. The bigger challenge is converting those strengths into routine systems: clear standards, regular supervision, maintenance, reporting, and follow up.

The main output is now a BLAM scoping report and a clearer pathway for future work. The pilot also helped me refine the project direction. I now see BLAM not only as a biosafety assessment tool, but as a platform that can later be extended to test whether fragile laboratory systems are ready to use AI safely in biological work. 

What are your next steps?

My next step is to use the pilot report to seek follow-on support for a focused scale-up phase. The immediate plan is to refine BLAM into a cleaner, easier-to-use tool with a scoring guide, indicators, and practical reporting templates. I also want to develop a small AI-biosecurity add-on module that asks a simple question: can public health laboratories safely use AI tools for tasks such as drafting SOPs, summarizing guidance, or preparing reports without exposing sensitive information or trusting incorrect advice?

The next phase would likely focus on the National Public Health Reference Laboratory in Monrovia and a small number of linked county laboratories. It would test BLAM more systematically, run a simple AI-biosecurity tabletop exercise, and produce practical outputs such as a safe-use guide, verification checklist, and escalation pathway for AI-related laboratory mistakes. Longer term, I want BLAM to become a Liberia-owned tool that can support better biosafety investment, partner coordination, and preparedness planning.

Is there anything others could help you with?

First, technical feedback from people with experience in biosafety, biosecurity, laboratory systems, or AI-biosecurity would be very useful. I especially need reviewers who can help make BLAM simpler, safer, and more credible before wider use.

Second, introductions to potential funders or collaborators interested in global health security, AI safety, pandemic preparedness, or laboratory strengthening would help move the work from seed pilot to scale up.

Third, I would appreciate practical support in turning the BLAM report into stronger public-facing materials: a short concept note, a clean tool/manual, and a funder-ready scale-up proposal. My PhD is currently unfunded, so support for protected time is also important if I am to keep pushing this work forward properly.

What progress have you made since your last update?

We re-analyzed the pilot data and quantified the headline effect: a small oracle budget raised the dishonest debater's win rate by +7.2pp (95% CI [4.6, 10.2]).

Before scaling up, we audited the pilot code and found two serious bugs in the oracle channel: every "NOT ADDRESSED" reply was miscoded to "NO", and ~100% of oracle queries were sent garbled. We retracted the mechanism conclusions and corrected the write-ups.

We then rebuilt the harness and launched a pre-registered validation run (in flight now, ~$150-200 of the grant): the same 318 transcripts re-judged under six arms, including a clean harness, a faithful bug replay, and a placebo oracle. Gates were frozen before any clean data existed. Pre-registration: https://github.com/jackmaiorino/selvarath-debate/blob/rerun-new-models/docs/rejudge-protocol.md and corrected report: https://github.com/jackmaiorino/selvarath-debate/blob/rerun-new-models/reports/2026-07-06-preliminary-findings.md

What are your next steps?

The pre-registered gates decide: if the effect survives the clean harness, we proceed to the proposed judge x debater capability grid (most of the grant). If the placebo explains it, we pivot to studying deliberation effects. If it collapses, we publish the artifact result and a decomposition of what each bug contributed. Write-up either way, including negative results.

Is there anything others could help you with?

Methods scrutiny of the pre-registered protocol before results land, and pointers to related work on oracle/verification interfaces or deliberation-length effects in LLM judging.

@JueYan Is it rude to ask you to take a look at my project? I'm bootstrapping since over a year and it's slowing things down by now. Thanks if you have/take the time to look, it's all I ask (apart from the funding ofc ;)

@nicholas Thank You SO very much!

That's a very generous offer, thank you @nicholas

I'll send you a mail tomorrow :)

@Lisa-Intel I think you missed the point haha, I'm providing the software to clients for no charge. I am accepting donations, though. E-mail me at nicholas@voltarecovery.org if you want to discuss this further.

Hi Nicholas, thanks!

I'd say my biggest bottleneck is building the core platform fast enough while staying focused. There are definitely areas where additional software could accelerate development, especially around AI orchestration, governance, integrations and developer tooling. Right now I'm keeping external spending very low while raising funding, but I'd love to stay in touch.

Coincidentally, I'm also exploring potential collaborations with researchers at the University of Florida (Department of Electrical and Computer Engineering) around autonomous AI systems, so it's nice to connect with you too.
@nicholas

Hey Pedro, out of curiosity, except for money, what is your biggest pain points that can be made into a software however it will cost a lot to create? I'm looking for clients honestly.

What i am trying to reach :

AATM started as a fundamental research project: can artificial cognition emerge from architecture, rather than from explicitly programmed cognitive functions?

The project has now found a major application axis: embodied intelligence for robotics.

One of the hardest problems in this field is not perception.
It is not movement.
It is not even learning.

It is internal regulation.

In living organisms, homeostasis is not a simple configuration file. The stability of internal states is the result of billions of years of evolution. Trying to manually find the perfect internal balance for one artificial organism is likely to be fragile, slow and structurally risky.

AATM changes the scaling strategy.

The goal is not to build one perfect artificial organism.

The goal is to build populations of imperfect blobs.

Each blob is a small autonomous sensorimotor unit: it has its own internal state, its own perception loop, its own movement, its own prediction errors, its own reinforcement dynamics and its own mini mammalian-inspired brain.

Some blobs may become too exploratory.
Some may become too conservative.
Some may collapse.
Some may become apathetic.
Some may discover useful behavioral patterns.

That is not a failure of the architecture.

That is the tuning surface.

Instead of fine-tuning one monolithic agent from the inside, AATM introduces a higher layer above the population. This layer can observe blobs over time, compare their behaviors, modify their parameters, duplicate useful configurations, remove unstable ones and consolidate what emerges across the swarm.

The higher layer does not need to have the same homeostasis as the blobs.

The blobs are driven by local regulation: action, energy, tension, prediction, novelty, saturation.

The upper layer is driven by population-level regulation: diversity, stability, coverage, robustness and collective performance.

This is the key shift.

AATM does not try to solve embodied intelligence by putting everything into one body and hoping the internal state remains stable.

It turns the problem into a distributed architecture:

local embodied agents below,
structural consolidation above.

This makes failure local instead of global.
It allows fine-tuning without restarting the whole system.
It creates a path from digital demonstrators to robotic swarms.
And it gives embodied AI a scalable architecture beyond the monolithic model.

What started as fundamental research into artificial cognition is becoming a practical path toward robotizable artificial organisms: imperfect by design, tunable by population, and scalable through collective consolidation.

Note: This grant was given as part of the experimental alignment microgrant program. Because this is a weird experimental grant program aiming to make lots of unusual bets, grantees are not allowed to use this grant as a credential when applying to any other grant, or job, or on their resume/linkedin/etc. This page exists because Manifund does not allow private grants. If you are a grantmaker or hiring manager, please disregard this grant when making decisions.

Maya, I appreciate the trust. Would you like to have a call? I also very much care about wild animals. Also happy to explain anything or keep you up to date on things